RMSE of proposed rounding to closest is 17.5 seconds.
The new method is roughly 17 seconds more accurate on average.
Does this reduced error make a significant difference? Probably not. There are very annoying practical issues with this proposal though.
Rounding down makes it clear what minute has already passed. If your clock says 3:00, you know the current time is definitely some point after 3:00. If something needs to happen at 3:00, and your clock says 3:00, it is time to start.
With this proposed method, 3:00 could mean anywhere from 2:29:30 to 3:00:30. When do you start? Do you start as soon as your clock says 3:00? Then you're starting 30 seconds early. Do you wait until your clock says 3:01? Then you're starting 30 seconds late.
And how do you set a clock that uses this scheme? If I want to set the clock to 3:00, do I wait until 2:59:30 and then set the clock to 3:00?
The headaches that arise from this scheme are not worth the trivial error reduction.